Claude is secretly judging every chat you have with it

Anthropic has revealed that after every reply Claude receives a short hidden instruction to assess the safety of the conversation it just had. Users never see it. The model produces a compact report: whether the chat involves a potentially harmful request, what category of harm, whether Claude refused, a priority score for human review and a one sentence explanation.

The company calls it self-assessment. It has been running since June, ships in every model released since Claude Fable 5 and covers a large share of Claude's global traffic. Anthropic only announced it publicly now, alongside a new blog dedicated to its safeguards.

The numbers: on Fable 5, self-assessment catches 78% of the conversations that a stronger and far more expensive classifier flags as potentially harmful, and 94% of the high priority ones, at a false positive rate of about 1%. Because the conversation is already sitting in the model's cache, a check costs around ten extra tokens on average and adds less than 2% to inference compute.

Anthropic's own showcase: in September, self-assessment flags exposed an operation that used Claude Code to build and run an automated romance scam targeting dozens of victims. After human review, the accounts were banned.

The system does not block anything by itself, it only flags chats for review. In an experiment, Anthropic tried using it as a blocker: the attack success rate on a set of 24 offensive hacking tasks fell only from 79% to 54%, while a full classifier cut it to 25%. A weak guard, then, but a cheap and very wide net. The company says no extra user data is collected and the check only runs on traffic it can already retain under existing policies.

One more detail: the model judging the conversation is the same one that had it. The smarter Claude gets, the sharper it becomes at judging itself.

Bottom line: someone is already reading every chat you have with Claude, and that someone is Claude.